An analysis of the RNN-based spoken term detection training

Abstract

This paper studies the training process of the recurrent neural networks used in the spoken term detection (STD) task. The method used in the paper employ two jointly trained Siamese networks using unsupervised data. The grapheme representation of a searched term and the phoneme realization of a putative hit are projected into the pronunciation embedding space using such networks. The score is estimated as relative distance of these embeddings. The paper studies the influence of different loss functions, amount of unsupervised data and the meta-parameters on the performance of the STD system.